You are on page 1of 6

The Martingale Stopping Theorem

Scott M. LaLonde
February 27, 2013

Abstract
We present a proof of the Martingale Stopping Theorem (also known as
Doobs Optional Stopping Theorem). We begin with some preliminaries on
measure-theoretic probability theory, which allows us to discuss the definition
and basic properties of martingales, We then state some auxiliary results and
use them to prove the main theorem.

Introduction

Recall that a martingale is (informally) a random process X = {Xn } which models


a players fortune in a fair game. That is to say, his expected fortune at time n given
the history up to this point is equal to his current fortune:
E(Xn |X1 , . . . , Xn1 ) = Xn1 .
This in turn implies that for all n,
E(Xn ) = E(Xn1 ) = = E(X1 ) = E(X0 ),
so the players expected fortune at any time is equal to his starting expected fortune.
It is natural to ask whether the game remains fair when stopped at a randomly
chosen time. Loosely speaking, if T is a random stopping time and XT denotes the
game stopped at this time, do we have
E(XT ) = E(X0 )
as well? In general the answer is no, as Doyle and Snell point out. They envision
a situation where the player is allowed to go into debt by any amount and to play
for an arbitrarily long time. In such a situation, the player will inevitably come out
ahead. There are conditions which will guarantee fairness, and Doyle and Snell [2]
give them in the following theorem, which is phrased in the context of gambling.
Theorem (Martingale Stopping Theorem). A fair game that is stopped at a random
time will remain fair to the end of the game if it is assumed that:
1

(a) There is a finite amount of money in the world.


(b) A player must stop if he wins all of this money or goes into debt by this amount.
Our goal is to develop a more formal statement of this theorem, called Doobs
Optional-Stopping Theorem, and then to prove it. We will start with some general
background material on probability theory, provide formal definitions of martingales
and stopping times, and finally state and prove the theorem. It should be noted that
our exposition will largely be based on that of Williams [4], though a nice overview
of martingales and various results about them can be found in Doob [1].

Preliminaries

Modern approaches to probability theory make much use of measure theory. Since
the proof of Doobs theorem will rely heavily on some sort of integral convergence
theorem (namely the Dominated Convergence Theorem), we need to introduce some
background that places probability theory within the realm of measure theory.
In modern probability theory the model for a random experiment is called a
probability space. This is a triple (, , P), where
is a set, called the sample space.
is a -algebra of subsets of .
P is a probability measure on (, ), i.e. every set in is measurable and
P() = 1.
The notion of a probability space generalizes ideas from discrete probability. We
have already mentioned that is the sample space of an experiment. The -algebra
represents the set of possible outcomes, or the events to which one can assign a
probability. The measure P gives the probability that an outcome occurs.
Of course in discrete probability one is usually interested in random variables,
which are real-valued functions on the sample space. For us, a random variable will
be a function X : R which is measurable with respect to . The expected value
of a random variable X is its integral with respect to the measure P:
Z
E(X) =
X() dP(),

and we will say that a random variable X is integrable if E(|X|) < . Finally, we
will need to make reference to the conditional expectation of a random variable: given
a sub--algebra A of , the conditional expectation E(X|A) is a random variable
which satisfies certain conditions related to X and A. The proper definition is quite
complicated, so one should simply think of E(X|A) as the expectation of X given
2

that the events contained in A have occurred. This description is not completely
accurate, but it should help the readers understanding of the uses of conditional
expectation in the sequel.

Martingales and Stopping Times

Now that we have the appropriate background material out of the way, we can formally define a martingale. Fix a probability space (, , P), and let X = {Xn }
n=0
be a sequence of random variables on .
Definition 1. A filtration on (, , P) is an increasing sequence F = {Fn }
n=1
F1 F2 F3
of sub--algebras of . The sequence {Xn }
n=1 is said to be adapted to F if Xn is
Fn -measurable for each n.
Remark 2. This definition may seem abstract, but it helps to keep the following
idea in mind. The -algebra Fn represents the information available to us at time n
in a random process, or the events that we can detect at time n. That the sequence
is increasing represents the fact that we gain information as the process goes on.
This idea can perhaps be made even more clear by pointing out that a common
choice for F is the natural filtration (or minimal filtration):
Fn = (X1 , . . . , Xn ).
In this case, Fn is the smallest -algebra on making the random variables X1 , . . . , Xn
measurable. The information available at time n is precisely that generated by the
Xi for 1 i n. Of course more information could be available, which would
correspond to a different choice of filtration F.
Filtrations are important because they provide a concise way of defining a martingale. With this in mind, let F = {Fn } be a fixed filtration on (, , P).
Definition 3. A random process X = {Xn } is called a martingale relative to F if
(a) X is adapted to F,
(b) E(|Xn |) < for all n, and
(c) E(Xn+1 |Fn ) = Xn almost surely.
As we have already discussed, we are interested in what happens when one stops
a martingale at a random time. To do this, we need a formal way of talking about
a rule for stopping a random process which does not depend on the future. This
leads to the following definition of a stopping time.
3

Definition 4. A map T : {1, 2, . . . , } is called a stopping time if


{T = n} = { : T () = n} Fn

(1)

for all n . We will say that T is almost surely finite if P({T = }) = 0.


Remark 5. Intuitively, T is a random variable taking positive integer values (and
possibly ) which gives a rule for stopping a random process. Condition (1) says
that the decision whether to stop or not at time n depends only on the information
available to us at time n (i.e. the history up to and including time n). No knowledge
of the future is required, since such a rule would surely result in an unfair game.
Let X = {Xn } be a random process, and let T be a stopping time. For any
positive integer n and any , we define
T n() = min{T (), n}.
With this notation, we can define a stopped process.
Definition 6. The stopped process X T = {XnT } is given by
XnT () = XT n() ().
A useful result that we will need for the proof of Doobs theorem (but that we
will not prove) says that X T inherits certain desirable properties from X.
Proposition 7. If X = {Xn } is a martingale, then the stopped process X T =
{XT n } is also a martingale. In particular, for all n we have
E(XT n ) = E(X0 ).
This is part (ii) of [4, Theorem 10.9], and an outline of the proof can be found
there. The proof is not difficult, but the details are not particularly enlightening
from our current perspective.

Doobs Optional-Stopping Theorem

We now have all the pieces in place to state and prove our main theorem. First we
need to formalize what it means to stop a process at a random time. Suppose
we have a martingale X = {Xn } and a stopping time T . Assume that T is almost
surely finite. Then we can define a random variable XT : R by
XT () = XT () (),
at least for outside some set of probability 0. (To make XT everywhere-defined,
we could set it equal to 0 on this null set.) Intuitively, E(XT ) represents the players
4

expected fortune when stopping at a random time. If we are to show that we still
have a fair game, we will need to check that
E(XT ) = E(X0 ).

(2)

Note that T n converges to T pointwise almost surely as n , so we have


XT n XT almost surely. Moreover, we know that E(XT n ) = E(X0 ) for all n. It
would be nice if we could conclude that E(XT n ) E(XT ), since we would then
have (2). This amounts to showing that
Z
Z
XT () dP()
XT n () dP()

as n . Convergence of this sort is by no means guaranteed. We need hypotheses


that will allow us to invoke convergence theorems from measure theory, with the
Dominated Convergence Theorem being the likely candidate.
We will prove the version of Doobs theorem given in [4, Theorem 10.10], which
is essentially the same as the formal statement given in class. The proof of part
(b) will differ slightly from Williams proof, however. In the process we will obtain
direct analogues of the Martingale Stopping Theorem from [2]. In this regard,
requirement that there is only a finite amount of money in the world can be
encoded by assuming that the random variables Xn are uniformly bounded; this is
condition (b) below. Similarly, the requirement that the player stop after a finite
amount of time is obtained by requiring that T be almost surely bounded, which is
condition (a). We also show that there is a third condition under which the theorem
holds; this condition is essentially limit on the size of a bet at any given time.
Theorem 8 (Doobs Optional-Stopping Theorem). Let (, , P) be a probability
space, F = {Fn } a filtration on , and X = {Xn } a martingale with respect to F.
Let T be a stopping time. Suppose that any one of the following conditions holds:
(a) There is a positive integer N such that T () N for all .
(b) There is a positive real number K such that
|Xn ()| < K
for all n and all , and T is almost surely finite.
(c) E(T ) < , and there is a positive real number K such that
|Xn () Xn1 ()| < K
for all n and all .
Then XT is integrable, and
E(XT ) = E(X0 ).
5

Proof. Note that in all three cases T is a.s. finite. By our previous discussion, this
implies that XT is a.s.-defined, and we have XT n XT almost surely. Furthermore, we know that XT n is integrable for all n, and that E(XT n ) = E(X0 ).
Suppose that (a) holds. Then for n N , we have T () n = T () for all .
Hence XT n = XT for n N , and it follows that XT is integrable with
E(XT ) = E(XT N ) = E(X0 ).
Now suppose that (b) holds. Then the boundedness condition on the Xn implies
that
|XT n ()| < K
for all n and all . Also, it is fairly easy to check that
T n()

XT n () = X0 () +

Xk () Xk1 ()

k=1

for all , so if (c) holds we have


T n()

|XT n ()| |X0 ()| +

|Xk () Xk1 ()| |X0 ()| + KT ().

k=1

Certainly X0 is integrable, and we have E(KT ) = KE(T ) < by assumption.


Therefore, in either case (b) or (c) we have bounded |XT n | by an integrable random
variable, so the Dominated Convergence Theorem applies. It follows that XT is
integrable, and
Z
Z
lim

XT n () dP() =

XT () dP().

Equivalently,
lim E(XT n ) = E(XT ).

But E(XT n ) = E(X0 ) for all n, so we have E(XT ) = E(X0 ), as desired.

References
[1] J.L. Doob, What is a martingale?, Amer. Math. Monthly 78 (1971), no. 5, 451
463.
[2] Peter G. Doyle and J. Laurie Snell, Random walks and electrical networks, Carus
Mathematical Monographs, Mathematical Association of America, Washington,
D.C., 1984.
[3] Efe
A.
Ok,
Probability
theory
with
https://files.nyu.edu/eo1/public/books.html.

economic

applications,

[4] David Williams, Probability with martingales, Cambridge University Press, Cambridge, 1991.